Subjects
The distribution of samples according to source, gene and cancer status is shown in Table
1.
Table 1
Characteristics of study subjects
EMBRACE | 247 | (64) | 92 | (42) | 0 |
kConFaB | 96 | (25) | 84 | (38) | 0 |
AJBCS | 19 | (5) | 22 | (10) | 0 |
ABCFS | 20 | (5) | 23 | (10) | 1 |
Total
|
382
| |
221
| | 1 |
Affected breast cancerb | 205 | (54) | 125 | (57) | 1 |
Affected ovarian cancerb | 24 | (6) | 8 | (4) | 0 |
Number of families | 257 | | 118 | | 1 |
A total of 604 living female Australian and British carriers of pathogenic
BRCA1 or
BRCA2 mutations were identified in 376 families from the following sources: the Epidemiological study of
BRCA1 and
BRCA2 Mutation Carriers (EMBRACE;
http://www.srl.cam.ac.uk/genepi/embraceindex.htm), the Kathleen Cuningham Consortium for Research into Familial Breast Cancer (kConFaB;
http://www.kconfab.org), the Australian Jewish Breast Cancer Study (AJBCS) [
13], and the Australian Breast Cancer Family Study (ABCFS) [
14,
15]. EMBRACE recruits participants from among women and men referred for genetic testing at clinical genetics centres in the UK and Eire. kConFaB recruits participants from multiple-case breast and ovarian cancer families referred for genetic testing at family cancer clinics in Australia and New Zealand. AJBCS recruits Ashkenazi Jewish women reporting a personal or family history of breast or ovarian cancer in a first- or second-degree relative, and living in Melbourne or Sydney, Australia. Finally, ABCFS is a population-based case–control-family study that includes women with a first primary breast cancer recruited through the Victorian and New South Wales cancer registries, and their affected and unaffected relatives. Apart from index cases recruited through cancer registries for the ABCFS, the cancer status of participants was based on self-report.
For samples recruited through EMBRACE, a pathogenic mutation was defined as an established disease-causing mutation under the classification scheme used by Breast Cancer Information Core
http://research.nhgri.nih.gov/bic/. For samples recruited through kConFaB, AJBCS and ABCFS, mutations were classified as pathogenic according to the criteria established by kConfab
http://www.kconfab.org/progress/classification.asp. Specifically, the criteria specify the following as being pathogenic: all truncating mutations, unless there is clear evidence that the mutation is a single nucleotide polymorphism (e.g. terminal
BRCA2 variant); and any variant that is well characterized in family studies of multiple generations, and not found in control individuals, that results in a nonconservative amino acid substitution, and occurs in a residue conserved across species and in a functional domain. All mutations included in the study that were shared across sites were classified as pathogenic according to both routes of definition.
Within Australia, ethical approvals were obtained from the ethics committees of the Peter MacCallum Cancer Institute, The Prince of Wales Hospital, The University of Melbourne, The Cancer Council New South Wales, The Cancer Council Victoria and the Queensland Institute of Medical Research. Ethical approval for the EMBRACE study was obtained from the Eastern Multicentre Research Ethics Committee and the relevant local ethics committees. Written informed consent was obtained from each participant.
Statistical methods
Individuals with a first diagnosis of primary invasive breast cancer were considered to be affected, whereas individuals with no reported breast or ovarian cancer were censored at age at interview. Individuals with a first diagnosis of primary ovarian cancer were censored as unaffected at age at onset of ovarian cancer, and selected analyses were also performed in which individuals with a first primary diagnosis of ovarian cancer were excluded. All individuals were censored at age of prophylactic mastectomy. Individuals reporting prophylactic surgery included a single BRCA2 carrier who was subsequently diagnosed with multiple breast cancers 4 and 5 years after surgery, and an additional 12 unaffected individuals (7 BRCA1 and 5 BRCA2 carriers) with surgery 1–11 years before interview (average 3 years). Prophylactic oophorectomy of affected and unaffected individuals was controlled for by adjustment as a time-dependent covariate, as described below.
Linear regression was used to assess the association of AR CAG repeat length (smaller allele size, larger allele size and average allele size) with potential confounders within the subset of 364 BRCA1 carriers and 209 BRCA2 carriers for whom information was available. The potential confounders included year of birth (categorized into subgroups 1910–1949, 1950–1959 and 1960–1979), age at menarche (categorized as ≤ 11, 11.5–12, 12.5–13, 13.5–14 or ≥ 14.5 years), oral contraceptive pill use (ever/never) and parity (categorized as 0 or ≥ 1 live births before censored age). Questionnaire information was available from participants on age at first and last live birth, but not age at each live birth. Hence, it was not possible to assess association with parity as an absolute number of live births before censored age, but rather only as a never/ever variable. Associations were assessed separately for affected and unaffected women.
The primary analyses of association between
AR genotype and disease risk were performed using Cox regression with time to breast cancer onset as the end-point.
AR CAG repeat length was defined as follows: a binary variable, defined by cut-points investigated in the hypothesis-generating study conducted by Rebbeck and coworkers [
7] (namely one or more allele of ≥ 28 CAG repeats, ≥ 29 CAG repeats, or ≥ 30 CAG repeats); or a continuous variable, using the length of the smaller of the two alleles (
AR small CAG), the larger of the two alleles (
AR large CAG), and the average length of a participant's two alleles (
AR average CAG). Rate ratios (RRs) and 95% CIs were estimated with adjustment for source group (as indicated in Table
1) and ethnicity (non-Jewish Caucasian, Jewish, other). Analyses were complicated by the fact that more than one mutation carrier could come from the same family and could not therefore be considered independent. Standard Cox regression provides unbiased RR estimates but their standard errors and CIs are incorrect. This was rectified by computing the confidence limits for the RRs using Huber's sandwich estimator of the covariance matrix [
17]. This allows for variation between carriers from the same family without modelling their dependence explicitly. Further analyses adjusted for oophorectomy and parity as time-dependent covariates, and for age at menarche, oral contraceptive pill use and year of birth. Oophorectomy before censored age at interview or diagnosis of breast cancer was reported by 39
BRCA1 carriers (10 with primary breast cancer) and 21
BRCA2 carriers (9 with primary breast cancer) with genotype information available.
RRs were estimated separately for BRCA1 and BRCA2 carriers, including in both analyses the single individual with a mutation in both genes. Models adjusting for only group and ethnicity included all individuals with genotype information, namely 376 BRCA1 carriers (with 200 events) and 219 BRCA2 carriers (with 122 events). The sample size for BRCA1 carriers with a putative risk allele was 28 (14 events) for the ≥ 28 CAG cut-point, 26 (13 events) for the ≥ 29 CAG cut-point, and 11 (4 events) for the ≥ 30 CAG cut-point. Similarly, for BRCA2 carriers it was 17 (10 events) for the ≥ 28 CAG cut-point, 14 (7 events) for the ≥ 29 CAG cutpoint, and 11 (5 events) for the ≥ 30 CAG cut-point. Full models adjusting for year of birth and additional hormonal variables included the 364 BRCA1 and 219 BRCA2 carriers with full information on potential confounders, comprising 193 and 116 events, respectively.
In addition, analyses were carried out separately for subgroups of
BRCA1 and
BRCA2 carriers defined by mutation position in relation to proposed AR-binding domains, and/or
in vitro data regarding mutation effect on AR transactivation [
10,
11]. For
BRCA1, subgroups were defined by the mutation position either relative to amino acid 1365 (< or ≥ nucleotide 4213), because mutations 5' of amino acid 1365 have been shown to have a markedly decreased effect on AR transactivation [
11]. This created subgroups including 314 and 62 individuals. In addition,
BRCA1 subgroups were defined by mutation relative to amino acid 1065 (< or ≥ nucleotide 3311), because this defines the 3' end of the
BRCA1 fragment shown
in vitro to bind the
AR amino-terminal domain containing the CAG-encoded polyglutamine tract [
11], creating subgroups of 210 and 166 individuals.
BRCA2 subgroups were defined by mutation relative to amino acid 1042 (< or ≥ nucleotide 3352), because it has been shown that the
BRCA2 L1042X mutation does not enhance AR transactivation [
10]. Subgroup sample sizes were 42 and 177. For both
BRCA1 and
BRCA2 subgroup analyses, the 3' and 5' subgroups were termed domain 1 and domain 2, respectively. Protein truncating and splice mutations were stratified into domain 1 or domain 2 according to their nucleotide/amino acid position, whereas all missense mutations were included in domain 2 because these nontruncating mutations may act in a dominant-negative manner.
Although the primary analysis provides a valid test of the association between a genotype and disease risk, it may not provide a consistent estimate of the RR because the disease status of the individuals may have affected the likelihood of ascertainment (for the non-population-based studies). Oversampling of affected individuals is likely, as is presentation of affected carriers at later mean age than unaffected carriers. To correct for this potential bias, we also conducted secondary analyses using the weighted Cox regression approach as described by Antoniou and coworkers (unpublished data), in which individuals are weighted such that the observed breast cancer incidence rates in the study sample are consistent with established breast cancer risk estimates for
BRCA1 and
BRCA2 mutation carriers. Antoniou and coworkers (unpublished data) have shown that this approach gives estimates that are close to unbiased, but with some loss of power as compared with the standard unweighted approach. Weights were computed separately for
BRCA1 and
BRCA2 mutation carriers using the breast cancer incidence rate estimates reported in the meta-analysis conducted by Antoniou
et al. [
18]. A global set of weights was computed because the number of mutation carriers by study was too small to compute reliable study-specific weights. Moreover, Antoniou
et al. [
18] found no significant differences in the
BRCA1 and
BRCA2 cancer risks by country or study centre. As for unweighted analyses, confidence limits for the risk ratio were calculated using a robust variance approach to allow for the dependence among individuals.
We evaluated the power of detecting the effects reported by Rebbeck and coworkers [
7] in our samples of
BRCA1 and
BRCA2 mutation carriers using simulations. For this purpose, we assumed the age distribution of the affected and unaffected carriers in our sample (Table
1) and simulated among them risk factors with risk ratios 1.8 and 2.6 and the frequencies for the ≥ 28 and ≥ 29 CAG cut-points observed in our sample. The data were then analyzed using unweighted Cox regression. We conducted 1000 simulations per model. More details about the simulations are available from the authors of the present report. The power of detecting risk ratios of 1.8 and 2.6 was estimated to be 51% and 92%, respectively, for the sample of
BRCA1 mutation carriers and 28% and 78% for the sample of
BRCA2 mutation carriers.
R version 1.9.0 (R Foundation for Statistical Computing, Vienna, Austria) was used to perform the unweighted Cox regression, and STATA version 7 (Stata Corporation, College Station, TX, USA) was used for the weighted analyses.